On Mon, Nov 28, 2016 at 11:25 PM, Andres Freund <andres@anarazel.de> wrote:
> On 2016-11-28 09:55:00 -0500, Robert Haas wrote:
>> I think we should go with this approach. I don't think it's a good
>> idea to have the heapam layer know about executor slots.
>
> I agree that that's not pretty.
>
>> Even though
>> it's a little sad to pass up an opportunity for a larger performance
>> improvement, this improvement is still quite good.
>
> It's imo not primarily about a larger performance improvement, but about
> avoid possible regressions. Doubling deforming of wide tuples will have
> noticeable impact on some workloads. I don't think we can disregard
> that.
I wasn't proposing to disregard that. I'm saying hoist the tests up
into nodeSeqScan.c where they can use the slot without breaking the
abstraction. It's not quite as fast but it's a lot cleaner.
--
Robert Haas
EnterpriseDB: http://www.enterprisedb.com
The Enterprise PostgreSQL Company